Integrating GAN to MFCC-Assisted RNN-LSTM Algorithms for Identifying Types of Tribes Based on English Pronunciation

Many studies based on Natural Language Processing (NLP) that have been carried out focus on the results of speech recognition on devices, but do not discuss the speech recognition process that is applied until the device works according to voice commands from the user. One of the factors for success in the classification process is the amount of unbalanced data. Imbalances in the data can impact the performance of the Deep Learning architecture used to create classifier models. Therefore, it is proposed to integrate the Generative Adversarial Network (GAN) algorithm and the Mel Frequency Cepstral Coefficient (MFCC) algorithm assisted by RNN-LSTM which is used in Speech Recognition in English Pronunciation by Non-Native Speakers. The proposed model is able to produce ethnic recognition based on the pronunciation made by a speaker. The study was conducted using 2,722 voice samples from 4 indigenous tribes in Indonesia, namely, the Ambonese, Sundanese, Javanese and Betawi. By using the proposed method and after testing, a classifier model was obtained with an accuracy of 78.12%.